Welcome back to Daily Zaps, your regularly-scheduled dose of AI news ⚡️ 

Here’s what we got for ya today:

• 🤖 Introducing ChatGPT Gov

• 🖥️ Meta’s plans to bring 1.3 million GPUs online this year

• 🏆 DeepSeek's approach emphasizes talent over secrecy

• 🖼️ DeepSeek releases image model Janus-Pro

Let’s get right into it!

Sign Up | Advertise | Tools

STARTUPS

Introducing ChatGPT Gov

OpenAI has launched ChatGPT Gov, a tailored version of its AI chatbot designed for U.S. government agencies, providing secure access to its advanced GPT-4o model for tasks such as text interpretation, coding, and translation. ChatGPT Gov allows agencies to deploy the tool in their own Microsoft Azure cloud environments, enabling them to manage security, privacy, and compliance requirements under frameworks like FedRAMP High.

Already in use by agencies like the Air Force Research Laboratory and Los Alamos National Laboratory, the tool supports scientific research, administrative tasks, and translation services. By offering features like Custom GPTs, conversation sharing, and an administrative console, OpenAI aims to enhance government efficiency, productivity, and innovation while safeguarding sensitive data.

BIG TECH

Meta’s plans to bring 1.3 million GPUs online this year

Meta is set to bring 1GW of compute power online this year and grow its fleet to over 1.3 million GPUs, backed by a $60-65 billion capital expenditure investment focused on data centers, servers, and expanding its AI teams. This represents a significant increase from last year's $35-40 billion capex. CEO Mark Zuckerberg highlighted a $10 billion, 2GW+ data center project in Richland Parish, Louisiana, which will be the company's largest facility to date, spanning four million square feet and expected to complete by 2030.

Emphasizing 2025 as a pivotal year for AI, Zuckerberg framed Meta's investments as essential to driving core products, fostering innovation, and maintaining American technology leadership, contrasting with OpenAI’s ambitious but uncertain $500 billion Stargate Project.

STARTUPS

DeepSeek's approach emphasizes talent over secrecy

DeepSeek's advancements in AI have sparked significant reactions, revealing a major shift in the global AI landscape. The company, leveraging innovations like mixture-of-experts (MoE), multi-head latent attention (MLA), and reinforcement learning without human feedback (RLHF), developed its V3 and R1 models with remarkable cost efficiency. V3 demonstrated exceptional optimization for H800 GPUs, overcoming U.S. chip bans, while R1 rivaled OpenAI's o1 in reasoning capabilities.

The breakthroughs suggest a world of cheaper training and inference, challenging Nvidia's dominance and unsettling markets. DeepSeek's open-weight approach emphasizes talent over secrecy, contrasting U.S. firms' push for regulation. This moment highlights China's software prowess and suggests a paradigm shift toward commoditized AI models. The U.S. faces a choice: compete through innovation or double down on restrictive policies. Regardless, consumers and businesses are poised to benefit most from increasingly accessible AI.

STARTUPS

DeepSeek releases image model Janus-Pro

Chinese startup DeepSeek has launched Janus-Pro, an AI image-generation model designed to compete with U.S. rivals like DALL-E 3 and Stable Diffusion, following the viral success of its R1 chatbot model. Janus-Pro, a multimodal model that generates images from text prompts, claims to outperform competitors in image quality and accuracy while leveraging a novel autoregressive framework to address challenges in visual analysis and generation.

The back-to-back launches highlight China's growing presence in the AI landscape, with DeepSeek utilizing less advanced Nvidia chips to achieve impressive results despite U.S. export restrictions. Janus-Pro is available for download on the AI platform Hugging Face.